deep-research

claude-skills

Production-grade Claude Code skills, auto-research validated.

Skills

Skill	Description	Validated	Lines
/deep-research (or /dr)	Deep Research — exhaustive, evidence-based investigation of any topic	89 runs, 85+ domains, 24.3/25 judge	187

Install

# As a Claude Code plugin (all skills)
/plugin install deep-research@yachi/claude-skills

# Or install manually
git clone https://github.com/yachi/claude-skills.git
cp -r claude-skills/skills/deep-research ~/.claude/commands/deep-research
# For short alias:
ln -s ~/.claude/commands/deep-research ~/.claude/commands/dr

/deep-research (alias: /dr)

Industrial-grade research skill. Every claim survives cross-examination by domain experts.

9 phases: Clarify → Decompose → Multi-Source Investigation → Industry Standards Audit → Quantitative Analysis → Adversarial Review → Practitioner Check → Verify → Synthesis

Standards: ICD 203 (9/9), GRADE, NIST reproducibility, CRAAP

Auto-research validated:

• 89 runs across 85+ domains (tech, finance, medical, legal, industrial, adversarial)
• 23 regex evals + LLM-as-judge (fact-checks 3 claims per output)
• Judge mean: 24.3/25, fact-check accuracy: 26/27
• Self-improving loop (Karpathy-inspired): evolve-v2.md

Includes:

• 98 research outputs across every domain imaginable
• 9 LLM judge assessments with fact-check details
• Original benchmark data (with-skill vs baseline, 5 domains)
• Visual benchmark dashboard
• Auto-research infrastructure (grade.py, judge.md, evolve loops)

See skills/deep-research/ for full docs, examples, and the auto-research toolkit.

Structure

claude-skills/
├── .claude-plugin/marketplace.json
├── skills/
│   └── deep-research/
│       ├── SKILL.md                    ← the skill (187 lines)
│       ├── eval/
│       │   ├── grade.py                ← 23-eval regex grading
│       │   └── judge.md                ← LLM fact-checker (/25)
│       ├── autoresearch/
│       │   ├── evolve-v2.md            ← self-improvement loop
│       │   ├── evolve-v1.md            ← legacy loop
│       │   ├── program.md             ← test prompts + criteria
│       │   ├── log.md                  ← full 89-run log
│       │   ├── outputs/               ← 89 runs + 9 judges
│       │   └── examples/              ← curated best/worst
│       └── benchmark/
│           ├── iteration-1/            ← with-skill vs baseline
│           └── dashboard/              ← visual dashboard
├── LICENSE (MIT)
└── README.md

Adding new skills

skills/
├── deep-research/  ← existing
│   └── SKILL.md
└── your-skill/  ← add here
    └── SKILL.md

Then add to .claude-plugin/marketplace.json.

License

MIT